2019/10/15 by Ali Mohammed, Ahmed Eleliemy, Mohammed, Ali +7
Computer Science · #Cloud Computing and Resource Management #Distributed #Distributed and Parallel Computing Systems #FOS: Computer and information sciences #Parallel #Parallel Computing and Optimization Techniques #Performance (cs.PF) #and Cluster Computing (cs.DC)
paper · pdf · doi:10.48550/arxiv.1910.06844
openalex publication_date 2019/10/15 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28
Scientific applications often contain large, computationally-intensive, and\nirregular parallel loops or tasks that exhibit stochastic characteristics.\nApplications may suffer from load imbalance during their execution on\nhigh-performance computing (HPC) systems due to such characteristics. Dynamic\nloop self-scheduling (DLS) techniques are instrumental in improving the\nperformance of scientific applications on HPC systems via load balancing.\nSelecting a DLS technique that results in the best performance for different\nproblems and system sizes requires a large number of exploratory experiments. A\ntheoretical model that can be used to predict the scheduling technique that\nyields the best performance for a given problem and system has not yet been\nidentified. Therefore, simulation is the most appropriate approach for\nconducting such exploratory experiments with reasonable costs. This work\ndevises an approach to realistically simulate computationally-intensive\nscientific applications that employ DLS and execute on HPC systems. Several\napproaches to represent the application tasks (or loop iterations) are compared\nto establish their influence on the simulative application performance. A novel\nsimulation strategy is introduced, which transforms a native application code\ninto a simulative code. The native and simulative performance of two\ncomputationally-intensive scientific applications are compared to evaluate the\nrealism of the proposed simulation approach. The comparison of the performance\ncharacteristics extracted from the native and simulative performance shows that\nthe proposed simulation approach fully captured most of the performance\ncharacteristics of interest. This work shows and establishes the importance of\nsimulations that realistically predict the performance of DLS techniques for\ndifferent applications and system configurations.\n